CAISI Adds Microsoft, Google, and xAI to Pre-Deployment Frontier Access — Still Voluntary
Commerce’s CAISI announces agreements with Microsoft, Google DeepMind, and xAI for early frontier-model access and national-security evaluations — expanding prior OpenAI/Anthropic deals. Guardrails often stripped for probing; 40+ evals completed. Institutionalized voluntary vetting, not a binding pre-release statute.
Washington is building a voluntary pre-deployment evaluation path, not an EU-style statute. CAISI’s new agreements expand who shares models early — Microsoft, Google DeepMind, and xAI joining the OpenAI/Anthropic precedent — while Mythos-class cyber capability keeps the urgency high.
CAISI at the U.S. Department of Commerce (May 5) announced agreements with Microsoft, Google DeepMind, and xAI for early access to new frontier AI models for national security and security-risk evaluations before deployment. Builds on existing deals with OpenAI and Anthropic; tied to July 2025 AI Action Plan pledges. CAISI (within NIST) has completed more than 40 evaluations of cutting-edge models.
Scope
- Developers provide early access, often with safety guardrails removed for probing.
- CAISI assesses capabilities, security risks, and national-security implications before public deployment.
- Collaborative research on capabilities and safeguards.
Microsoft: adversarial assessments; co-develop systematic evaluation approaches, shared frameworks/datasets/workflows; parallel UK AISI agreement. Google DeepMind and xAI: models for similar capability/risk evaluations. CAISI Director Chris Fall: independent measurement science is essential for frontier AI national-security implications.
Claims vs checks
Agreements and Microsoft’s own On the Issues post are primary; Reuters confirms Microsoft/Google/xAI early-access framing. “Before deployment” is partnership practice — not a legal prohibition on shipping without CAISI sign-off. Mythos dual-use alarm is contemporaneous context, not a CAISI finding in this release.
Limits
- Voluntary; non-participants and open-weight paths sit outside.
- Eval results largely non-public.
- Guardrail-stripped access is for evaluators — residual leakage risk is a process question.
Sources
- Reuters: “Microsoft, Google and xAI to give US government early access to AI models for security checks” (May 5, 2026).
- Microsoft: “Advancing AI evaluation with CAISI (US) and AISI (UK)” (May 5, 2026).
- CAISI/NIST statements via Commerce reporting.